Agentic Coding: Myths and Realities of an Announced Revolution

⚡

Key Takeaways

1Agentic AI promises rapid coding but raises challenges in security and maintenance.

2Exaggerated narratives about agentic AI obscure the real issues of control and code quality.

3Automated testing is not enough to ensure the robustness of AI-generated software in real-world conditions.

💡Why it matters — Developers need to adapt their practices to effectively integrate AI without compromising software quality.

The Promise of Agentic AI

Agentic AI is often touted as a miracle solution for accelerating software development. However, it brings with it hidden challenges related to testing, security, and maintenance. To prevent these issues from jeopardizing projects, developers must rethink their methods of management, validation, and oversight of machine-generated software at scale.

Myths and Realities of Agentic Coding

Two narratives dominate the discourse on agentic coding. The first claims that a simple phrase can enable AI to create a million-dollar application. The second asserts that AI, by writing all the code, leaves humans completely unaware of its content, leading to inevitable failure. These narratives oversimplify a more complex reality. This article examines the maintenance and sustainability issues that arise when control of coding is handed over to a machine.

1. The Myth of Losing Control

Engineering managers have always had to supervise contractors. Assigning tasks and evaluating the work produced is essential in software engineering. While the pessimism surrounding agentic coding is often exaggerated, it contains a kernel of truth. Without quality standards and practices, the code can become problematic.

It is advisable to provide the AI with detailed specification documents, but misinterpretation can lead to deviations.
Preferring to give the AI simple tasks helps reduce the risks of losing sight of the overall plan.

2. The Myth of Real-World Readiness

A striking example of the fragility of automated testing systems is a friend who, despite the rigor of his tests, always manages to break the code. This can be attributed to a lack of internal knowledge about how the code functions.

Automated tests can identify if a recent fix has caused a problem, but they often do not cover edge cases.
Testing environments do not always reflect real-world conditions, which can mask significant issues.

To overcome this problem, it is crucial to test as an external user. Integrating adversarial testing practices and planning for edge case and error scenarios in every test plan is essential.

3. The Myth of Legacy Code

Over the years in the software industry, I have acquired rights to software intellectual properties. Most of the products I launched were initially coded by others. While this has its advantages, it also presents challenges.

Acquired products may carry technical debt and issues that make the software less valuable.
Often, the original developers no longer wish to take responsibility for maintenance and support.

Every experience of coding with AI will be akin to buying a house without a prior inspection, discovering hidden problems once the code is in production.